I have been building MetricSync, an iPhone AI food logging app, and one UX lesson keeps getting clearer:
The AI part is not enough.
A lot of AI food logging demos look good because the first step is impressive. Take a photo, scan a barcode, or type a messy description, then get a structured food log back.
That is useful, but it is only half the product.
The real product is what happens after the guess.
The first guess is usually not the final answer
Food is messy in a way that software demos usually hide.
A photo might show rice and chicken, but miss the sauce.
A barcode might identify the packaged item, but the serving size still needs attention.
A text entry might say "burrito bowl", but the actual meal depends on what went into it.
If the app treats the AI result as final, the user has to fight the product whenever it is slightly wrong.
That is where trust drops.
Not because the AI failed completely, but because the correction path feels expensive.
The review screen matters more than the prompt
For this kind of app, I think the review screen is a first-class feature.
The user should be able to quickly answer questions like:
- What did the app think I ate?
- What serving size did it choose?
- Which items can I edit quickly?
- Can I fix one part without restarting the whole log?
- Can I save and move on without turning lunch into a data-entry chore?
That is less flashy than the AI demo, but it is probably more important for retention.
If correction feels normal, the user can forgive an imperfect guess.
If correction feels like starting over, the product loses the moment the demo gets a real meal wrong.
Multiple inputs need one correction loop
This is also why I do not think photo, barcode, and text should feel like three separate products.
They are three capture methods for the same job.
The output still needs to land in a clear, editable food log.
For MetricSync, the goal is:
- Capture the meal quickly
- Let AI do the first pass
- Show the result clearly
- Make correction cheap
- Let the user move on
The correction loop is the product moat, not just the model call.
My current rule
When I look at an AI feature now, I ask:
How annoying is it when the model is 80 percent right?
That question catches a lot of product problems.
An 80 percent correct AI result can either feel magical or frustrating depending on the edit flow around it.
For food logging especially, the app needs to assume imperfect inputs, imperfect photos, and imperfect guesses.
That is the direction I am taking with MetricSync: iPhone food logging from photo, barcode, or text, with a fast review and correction flow instead of treating the first AI result as sacred.
MetricSync has a 3-day free trial, then it is $5/month: https://metricsync.download
Top comments (0)